[NeurIPS 2021 Spotlight] HELP: Hardware-adaptive Efficient Latency Prediction for NAS via Meta-Learning [Paper]

This is Official PyTorch implementation for HELP: Hardware-adaptive Efficient Latency Prediction for NAS via Meta-Learning.

@inproceedings{lee2021help,
    title     = {HELP: Hardware-Adaptive Efficient Latency Prediction for NAS via Meta-Learning},
    author    = {Lee, Hayeon and Lee, Sewoong and Chong, Song and Hwang, Sung Ju},
    booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
    year      = {2021}
}

Overview

For deployment, neural architecture search should be hardware-aware, in order to satisfy the device-specific constraints (e.g., memory usage, latency and energy consumption) and enhance the model efficiency. Existing methods on hardware-aware NAS collect a large number of samples (e.g., accuracy and latency) from a target device, either builds a lookup table or a latency estimator. However, such approach is impractical in real-world scenarios as there exist numerous devices with different hardware specifications, and collecting samples from such a large number of devices will require prohibitive computational and monetary cost. To overcome such limitations, we propose Hardware-adaptive Efficient Latency Predictor (HELP), which formulates the device-specific latency estimation problem as a meta-learning problem, such that we can estimate the latency of a model's performance for a given task on an unseen device with a few samples. To this end, we introduce novel hardware embeddings to embed any devices considering them as black-box functions that output latencies, and meta-learn the hardware-adaptive latency predictor in a device-dependent manner, using the hardware embeddings. We validate the proposed HELP for its latency estimation performance on unseen platforms, on which it achieves high estimation performance with as few as 10 measurement samples, outperforming all relevant baselines. We also validate end-to-end NAS frameworks using HELP against ones without it, and show that it largely reduces the total time cost of the base NAS method, in latency-constrained settings.

Prerequisites

Python 3.8 (Anaconda)
PyTorch 1.8.1
CUDA 10.2

Hardware spec used for meta-training the proposed HELP model

GPU: A single Nvidia GeForce RTX 2080Ti
CPU: Intel(R) Xeon(R) Silver 4114 CPU @ 2.20GHz

Installation

$ conda create --name help python=3.8
$ conda activate help
$ conda install pytorch==1.8.1 torchvision cudatoolkit=10.2 -c pytorch
$ pip install nas-bench-201
$ pip install tqdm
$ conda install scipy
$ conda install pyyaml
$ conda install tensorboard

1. Experiments on NAS-Bench-201 Search Space

2. Experiments on FBNet Search Space

3. Experiments on OFA Search Space

4. Experiments on HAT Search Space

1. Reproduce Main Results on NAS-Bench-201 Search Space

We provide the code to reproduce the main results on NAS-Bench-201 search space as follows:

Computing architecture ranking correlation between latencies estimated by HELP and true measured latencies on unseen devices (Table 3).
Latency-constrained NAS Results with MetaD2A + HELP on unseen devices (Table 4).
Meta-Training HELP model.

1.1. Data Preparation and Model Checkpoint

We include all required datasets and checkpoints in this github repository.

1.2. [Meta-Test] Architecture ranking correlation

You can compute architecture ranking correlation between latencies estimated by HELP and true measured latencies on unseen devices on NAS-Bench-201 search space (Table 3):

$ python main.py --search_space nasbench201 \
		 --mode 'meta-test' \
		 --num_samples 10 \
		 --num_meta_train_sample 900 \
                 --load_path [Path of Checkpoint File] \
		 --meta_train_devices '1080ti_1,1080ti_32,1080ti_256,silver_4114,silver_4210r,samsung_a50,pixel3,essential_ph_1,samsung_s7' \
		 --meta_valid_devices 'titanx_1,titanx_32,titanx_256,gold_6240' \                 
                 --meta_test_devices 'titan_rtx_256,gold_6226,fpga,pixel2,raspi4,eyeriss'

You can use checkpoint file provided by this git repository ./data/nasbench201/checkpoint/help_max_corr.pt as follows:

$ python main.py --search_space nasbench201 \
		 --mode 'meta-test' \
		 --num_samples 10 \
		 --num_meta_train_sample 900 \
                 --load_path './data/nasbench201/checkpoint/help_max_corr.pt' \
		 --meta_train_devices '1080ti_1,1080ti_32,1080ti_256,silver_4114,silver_4210r,samsung_a50,pixel3,essential_ph_1,samsung_s7' \
		 --meta_valid_devices 'titanx_1,titanx_32,titanx_256,gold_6240' \                 
                 --meta_test_devices 'titan_rtx_256,gold_6226,fpga,pixel2,raspi4,eyeriss'

or you can use provided script:

$ bash script/run_meta_test_nasbench201.sh [GPU_NUM]

Architecture Ranking Correlation Results (Table 3)

Method	# of Training Samples From Target Device	Desktop GPU (Titan RTX Batch 256)	Desktop CPU (Intel Gold 6226)	Mobile Pixel2	Raspi4	ASIC	FPGA	Mean
FLOPS	-	0.950	0.826	0.765	0.846	0.437	0.900	0.787
Layer-wise Predictor	-	0.667	0.866	-	-	-	-	0.767
BRP-NAS	900	0.814	0.796	0.666	0.847	0.811	0.801	0.789
BRP-NAS (+extra samples)	3200	0.822	0.805	0.693	0.853	0.830	0.828	0.805
HELP (Ours)	10	0.987	0.989	0.802	0.890	0.940	0.985	0.932

1.3. [Meta-Test] Efficient Latency-constrained NAS combined with MetaD2A

You can reproduce latency-constrained NAS results with MetaD2A + HELP on unseen devices on NAS-Bench-201 search space (Table 4):

$ python main.py --search_space nasbench201 --mode 'nas' \
                 --load_path [Path of Checkpoint File] \
                 --sampled_arch_path 'data/nasbench201/arch_generated_by_metad2a.txt' \
                 --nas_target_device [Device] \ 
                 --latency_constraint [Latency Constraint]

For example, if you use checkpoint file provided by this git repository, then path of checkpoint file is ./data/nasbench201/checkpoint/help_max_corr.pt, if you set target device as CPU Intel Gold 6226 (gold_6226) with batch size 256 and target latency constraint as 11.0 (ms), command is as follows:

$ python main.py --search_space nasbench201 --mode 'nas' \
                 --load_path './data/nasbench201/checkpoint/help_max_corr.pt' \
                 --sampled_arch_path 'data/nasbench201/arch_generated_by_metad2a.txt' \
                 --nas_target_device gold_6226 \ 
                 --latency_constraint 11.0

or you can use provided script:

$ bash script/run_nas_metad2a.sh [GPU_NUM]

Efficient Latency-constrained NAS Results (Table 4)

Device	# of Training Samples from Target Device	Latency Constraint (ms)	Latency (ms)	Accuracy (%)	Neural Architecture Config
GPU Titan RTX (Batch 256) `titan_rtx_256`	10	18.0 21.0 25.0	17.8 18.9 24.2	69.7 71.5 71.8	link link link
CPU Intel Gold 6226 `gold_6226`	10	8.0 11.0 14.0	8.0 10.7 14.3	67.3 70.2 72.1	link link link
Mobile Pixel2 `pixel2`	10	14.0 18.0 22.0	13.0 19.0 25.0	69.7 71.8 73.2	link link link
ASIC-Eyeriss `eyeriss`	10	5.0 7.0 9.0	3.9 5.1 9.1	71.5 71.8 73.5	link link link
FPGA `fpga`	10	4.0 5.0 6.0	3.8 4.7 7.4	70.2 71.8 73.5	link link link

1.4. Meta-Training HELP model

Note that this process is performed only once for all NAS results.

$ python main.py --search_space nasbench201 \
                 --mode 'meta-train' \
                 --num_samples 10 \
                 --num_meta_train_sample 900 \
                 --meta_train_devices '1080ti_1,1080ti_32,1080ti_256,silver_4114,silver_4210r,samsung_a50,pixel3,essential_ph_1,samsung_s7' \
                 --meta_valid_devices 'titanx_1,titanx_32,titanx_256,gold_6240' \           
                 --meta_test_devices 'titan_rtx_256,gold_6226,fpga,pixel2,raspi4,eyeriss' \
                 --exp_name [EXP_NAME] \
                 --seed 3 # e.g.) 1, 2, 3

or you can use provided script:

$ bash script/run_meta_training_nasbench201.sh [GPU_NUM]

The results (checkpoint file, log file etc) are saved in

./results/nasbench201/[EXP_NAME]

2. Reproduce Main Results on FBNet Search Space

We provide the code to reproduce the main results on FBNet search space as follows:

Computing architecture ranking correlation between latencies estimated by HELP and true measured latencies on unseen devices (Table 2).
Meta-Training HELP model.

2.1. Data Preparation and Model Checkpoint

We include all required datasets and checkpoints in this github repository.

2.2. [Meta-Test] Architecture ranking correlation

You can compute architecture ranking correlation between latencies estimated by HELP and true measured latencies on unseen devices on FBNet search space (Table 2):

$ python main.py --search_space fbnet \
	--mode 'meta-test' \
	--num_samples 10 \
	--num_episodes 4000 \
	--num_meta_train_sample 4000 \
	--load_path './data/fbnet/checkpoint/help_max_corr.pt' \
	--meta_train_devices '1080ti_1,1080ti_32,1080ti_64,silver_4114,silver_4210r,samsung_a50,pixel3,essential_ph_1,samsung_s7' \
	--meta_valid_devices 'titanx_1,titanx_32,titanx_64,gold_6240' \
	--meta_test_devices 'fpga,raspi4,eyeriss'

or you can use provided script:

$ bash script/run_meta_test_fbnet.sh [GPU_NUM]

Architecture Ranking Correlation Results (Table 2)

Method	Raspi4	ASIC	FPGA	Mean
MAML	0.718	0.763	0.727	0.736
Meta-SGD	0.821	0.822	0.776	0.806
HELP (Ours)	0.887	0.943	0.892	0.910

2.3. Meta-Training HELP model

Note that this process is performed only once for all results.

$ python main.py --search_space fbnet \
	--mode 'meta-train' \
	--num_samples 10 \
	--num_episodes 4000 \
	--num_meta_train_sample 4000 \
	--exp_name [EXP_NAME] \
	--meta_train_devices '1080ti_1,1080ti_32,1080ti_64,silver_4114,silver_4210r,samsung_a50,pixel3,essential_ph_1,samsung_s7' \
	--meta_valid_devices 'titanx_1,titanx_32,titanx_64,gold_6240' \
	--meta_test_devices 'fpga,raspi4,eyeriss' \
	--seed 3 # e.g.) 1, 2, 3

or you can use provided script:

$ bash script/run_meta_training_fbnet.sh [GPU_NUM]

The results (checkpoint file, log file etc) are saved in

./results/fbnet/[EXP_NAME]

3. Reproduce Main Results on OFA Search Space

We provide the code to reproduce the main results on OFA search space as follows:

Latency-constrained NAS Results with accuracy predictor of OFA + HELP on unseen devices (Table 5).
Validating obatined neural architecture on ImageNet-1K.
Meta-Training HELP model.

3.1. Data Preparation and Model Checkpoint

We include required datasets except ImageNet-1K, and checkpoints in this github repository. To validate obatined neural architecture on ImageNet-1K, you should download ImageNet-1K (2012 ver.)

3.2. [Meta-Test] Efficient Latency-constrained NAS combined with accuracy predictor of OFA

You can reproduce latency-constrained NAS results with OFA + HELP on unseen devices on OFA search space (Table 5):

python main.py \
	--search_space ofa \
	--mode nas \
	--num_samples 10 \
	--seed 3 \
	--num_meta_train_sample 4000 \
	--load_path './data/ofa/checkpoint/help_max_corr.pt' \
	--nas_target_device [DEVICE_NAME] \
	--latency_constraint [LATENCY_CONSTRAINT] \
	--exp_name 'nas' \
	--meta_train_devices '2080ti_1,2080ti_32,2080ti_64,titan_xp_1,titan_xp_32,titan_xp_64,v100_1,v100_32,v100_64' \
	--meta_valid_devices 'titan_rtx_1,titan_rtx_32' \
	--meta_test_devices 'titan_rtx_64'

For example,

$ python main.py \
	--search_space ofa \
	--mode nas \
	--num_samples 10 \
	--seed 3 \
	--num_meta_train_sample 4000 \
	--load_path './data/ofa/checkpoint/help_max_corr.pt' \
	--nas_target_device titan_rtx_64 \
	--latency_constraint 20 \
	--exp_name 'nas' \
	--meta_train_devices '2080ti_1,2080ti_32,2080ti_64,titan_xp_1,titan_xp_32,titan_xp_64,v100_1,v100_32,v100_64' \
	--meta_valid_devices 'titan_rtx_1,titan_rtx_32' \
	--meta_test_devices 'titan_rtx_64'

or you can use provided script:

$ bash script/run_nas_ofa.sh [GPU_NUM]

Efficient Latency-constrained NAS Results (Table 5)

Device	Sample from Target Device	Latency Constraint (ms)	Latency (ms)	Accuracy (%)	Architecture config
GPU Titan RTX (Batch 64)	10	20 23 28	20.3 23.1 28.6	76.0 76.8 77.9	link link link
CPU Intel Gold 6226	20	170 190	147 171	77.6 78.1	link link
Jetson AGX Xavier	10	65 70	67.4 76.4	75.9 76.4	link link

3.3. Validating obtained neural architecture on ImageNet-1K

$ python validate_imagenet.py \
		--config_path [Path of neural architecture config file]
		--imagenet_save_path [Path of ImageNet 1k]

for example,

$ python validate_imagenet.py \
		--config_path 'data/ofa/architecture_config/gpu_titan_rtx_64/latency_28.6ms_accuracy_77.9.json' \
		--imagenet_save_path './ILSVRC2012'

3.4. Meta-training HELP model

Note that this process is performed only once for all results.

$ python main.py --search_space ofa \
		--mode 'meta-train' \
		--num_samples 10 \
		--num_meta_train_sample 4000 \
		--exp_name [EXP_NAME] \
                --meta_train_devices '2080ti_1,2080ti_32,2080ti_64,titan_xp_1,titan_xp_32,titan_xp_64,v100_1,v100_32,v100_64' \
                --meta_valid_devices 'titan_rtx_1,titan_rtx_32' \
                --meta_test_devices 'titan_rtx_64' \
		--seed 3 # e.g.) 1, 2, 3

or you can use provided script:

$ bash script/run_meta_training_ofa.sh [GPU_NUM]

4. Main Results on HAT Search Space

We provide the neural architecture configurations to reproduce the results of machine translation (WMT'14 En-De Task) on HAT search space.

Efficient Latency-constrained NAS Results

Task	Device	Samples from Target Device	Latency	BLEU score	Architecture Config
WMT'14 En-De	GPU NVIDIA Titan RTX	10	74.0ms 106.5ms	27.19 27.44	link link
WMT'14 En-De	CPU Intel Xeon Gold 6240	10	159.6ms 343.2ms	27.20 27.52	link link

You can test models by BLEU score and Computing Latency.

Reference

Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks (ICML17)

Meta-SGD: Learning to Learn Quickly for Few-Shot Learning

Once-for-All: Train One Network and Specialize it for Efficient Deployment (ICLR20)

NAS-Bench-201: Extending the Scope of Reproducible Neural Architecture Search (ICLR20)

BRP-NAS: Prediction-based NAS using GCNs (NeurIPS20)

HAT: Hardware Aware Transformers for Efficient Natural Language Processing (ACL20)

Rapid Neural Architecture Search by Learning to Generate Graphs from Datasets (ICLR21)

HW-NAS-Bench: Hardware-Aware Neural Architecture Search Benchmark (ICLR21)

Official PyTorch Implementation of HELP: Hardware-adaptive Efficient Latency Prediction for NAS via Meta-Learning (NeurIPS 2021 Spotlight)

Related tags

Overview

[NeurIPS 2021 Spotlight] HELP: Hardware-adaptive Efficient Latency Prediction for NAS via Meta-Learning [Paper]

Overview

Prerequisites

Installation

Contents

1. Reproduce Main Results on NAS-Bench-201 Search Space

1.1. Data Preparation and Model Checkpoint

1.2. [Meta-Test] Architecture ranking correlation

1.3. [Meta-Test] Efficient Latency-constrained NAS combined with MetaD2A

1.4. Meta-Training HELP model

2. Reproduce Main Results on FBNet Search Space

2.1. Data Preparation and Model Checkpoint

2.2. [Meta-Test] Architecture ranking correlation

2.3. Meta-Training HELP model

3. Reproduce Main Results on OFA Search Space

3.1. Data Preparation and Model Checkpoint

3.2. [Meta-Test] Efficient Latency-constrained NAS combined with accuracy predictor of OFA

3.3. Validating obtained neural architecture on ImageNet-1K

3.4. Meta-training HELP model

4. Main Results on HAT Search Space

Reference

Owner

PyTorch implementation of Octave Convolution with pre-trained Oct-ResNet and Oct-MobileNet models

A hybrid SOTA solution of LiDAR panoptic segmentation with C++ implementations of point cloud clustering algorithms. ICCV21, Workshop on Traditional Computer Vision in the Age of Deep Learning

VOS: Learning What You Don’t Know by Virtual Outlier Synthesis

Official repository for ABC-GAN

Sentinel-1 vessel detection model used in the xView3 challenge

List of awesome things around semantic segmentation 🎉

User-friendly bulk RNAseq deconvolution using simulated annealing

Place holder for HOPE: a human-centric and task-oriented MT evaluation framework using professional post-editing

DeepFaceLive - Live Deep Fake in python, Real-time face swap for PC streaming or video calls

Companion repo of the UCC 2021 paper "Predictive Auto-scaling with OpenStack Monasca"

Exploit Camera Raw Data for Video Super-Resolution via Hidden Markov Model Inference

This repository contains implementations and illustrative code to accompany DeepMind publications

DCGAN LSGAN WGAN-GP DRAGAN PyTorch

An implementation of shampoo

Harmonic Memory Networks for Graph Completion

Deep Learning ❤️ OneFlow

A light and fast one class detection framework for edge devices. We provide face detector, head detector, pedestrian detector, vehicle detector......

ICON: Implicit Clothed humans Obtained from Normals (CVPR 2022)

Unified MultiWOZ evaluation scripts for the context-to-response task.

Open Source Light Field Toolbox for Super-Resolution